Replicating ‘How Newspapers Reveal Political Power’

Adapted from Ban et al. (2019)

Sean Hambali

February 21, 2024

Introduction

Motivation

How to measure power?
- Difficult to empirically observe power
Ban et al. (2019) provides an empirical measure of power via newspaper coverage
Main thesis: relative amount of space dedicated to a particular subject measures the relative power of that subject. Why?
Newsworthiness!
- Size/impact of story matters
- Prominence of actors also matters

Methods

What kind of power?

This paper specifically measures whether, and to what extent, political actors and offices possess the resources to influence political process.
- Provides cardinal measure of power
- Caveat 1: Newsworthiness does not necessarily translate to power, e.g. entertainment
- Caveat 2: Potential bias from news outlets’ ideological preferences
- Caveat 3: Does not measure the power of “shadowy” actors

Data

Primarily sourced from Newspapers.com
Stratified sample from 50% of archive
Focuses on years 1877-1977
Uses OCR tools and regex to count occurrences of political actors and offices in newspapers texts

(Validation) Results

Validation 1: Valuing congressional committees

Hypothesis: More powerful congressional committees should get more news coverage.
Ban et al. (2019) validates this hypothesis with committee “value” rankings constructed by Groseclose and Stewart (1998)

\[ \text{Relative Coverage of Committee}_i = \frac{\text{Committee}_i}{\sum^{19}_{i=1} \text{Committee}_i} \]

Result 1: Coverage vs Groseclose-Stewart rankings

Validation 2: Newspaper coverage of House Speakers

Hypothesis: There should be spike in coverage in the periods where senators serve as party leaders

Result 2: Coverage of speakers before, during and after leadership

Validation 3: Strong vs. weak mayors

Strong mayors (mayor-council):
- Mayor and city council are elected separately
- Mayor has broad executive authority
Weak mayors (council-manager):
- City council is the only elected office
- Council appoints city manager to oversee executive branch
- Mayor has little to no executive authority
Hypothesis: Weaker mayor authority upon transition to council-manager setting, thus lower mentions in newspapers.

\[ \hat{Y}_{ijt} = \frac{N_{ijt}}{\sum^{J}_{j=1} N_{ijt}} \]

where \(j \in \{\text{Mayor}, \text{City Manager}, \text{City Council}\}\), \(N\), and \(\hat{Y}\) denote mentions and relative proportions, respectively.

Result 3A: Coverage of city offices, before and after reforms

Result 3B: Effect of switching from Mayor-Council to Council-Manager

Original

Result 3B: Effect of switching from Mayor-Council to Council-Manager

Replication

Validation 4: Coverage of Massachussets Executive Council

Massachussets Executive Council is comprised of eight individuals elected from districts
In late 1964, the council was stripped of its statutory powers following corruption and bribery charges
Hypothesis: lower newspapers coverage on the Executive Council following the reform

\[ \text{EC}_t = \frac{\text{EC}_t}{\text{EC}_t + \text{G}_t} \]

where \(\text{EC}\) and \(\text{G}\) denote mentions of the Executive Council and the Governor in all available Massachussets newspapers, respectively.

Result 4: Coverage of Massachussets Executive Council

Application: Measuring the power of state party committees

Prediction: State committees are no longer as powerful as they once were
Ban et al. (2019) uses newspaper coverage to measure the power of state committees relative to local committees

\[ \text{Relative Party Mentions}_{it} = \frac{\text{Party}_{it}}{\text{Party}_{it} + \text{Candidates}_{it}} \]

Result 5: Party Committee Power in 9 U.S. States

A Closer Look at the Replication Exercise

Differences from the original results

Able to replicate all of the main findings
Minor differences in standard errors reported in regression table 3B

Replication diagnosis

The authors could not provide access to the raw newspapers data since they are proprietary data.
They only provide cleaned, aggregated term counts across years, states, etc.
As such, I wasn’t able to replicate the crucial first half of the paper, i.e., pre-processing the text.

Replication diagnosis

The authors used Stata to pre-process the term counts to generate measures of interest and R to visualize the output
I convert all Stata codes to R, posted all codes and output images/tables on GitHub.
I find one minor error in their code:

... %>%
  mutate(
    # Generate relative proportions
    rel_mayor_council_total = r_city_council + r_mayor,
    rel_mayor_council_control_total = r_control_sum_city_council + r_control_sum_mayor,
    rel_mayor_council = r_mayor / rel_mayor_council_total,
    rel_mayor_council_control = r_control_sum_mayor / rel_mayor_council_control_total, 
    
    # In the original paper, r_mayor was used instead of r_mayor_x
    rel_mayor_council_total_x = r_city_council_x + r_mayor_x, 
    rel_mayor_council_x = r_mayor_x / rel_mayor_council_total_x
  ) %>%
...

Possible extensions

Possible extensions:
- Application to the Indonesian case: rapid decentralization in 1998 shifts political power from central government to local government
- Possible extension to other mediums, e.g., news broadcast?
- Robustness checks to the DiD regression under new approaches by Callaway & Sant’Anna (2020), Chaisemartin & D’Haultfœuille (2020)
Limitations for extension:
- Limited nuance of the power being measured
- Limited use in developing country context due to pervasive power of “shadowy” political actors
- Limited use of newspapers in modern context, and possible changes to the incentive structure of newspaper outlets